concierge
I’ve just launched an AI assistant called concierge. For now, it lives on my website and can answer questions about me.
I generated the initial build with v0 and finished up in Cursor. For correctness, I augmented the LLM with basic memory using a JSON file that gets loaded into the system prompt.
You can try it here: concierge.opemipo.com.
My primary objective is to automate admin work. Eventually, concierge will be able to book my flights or renew this domain.
I’m also thinking about what it means to run an agent. This includes what it knows, how it feels, where it lives, what it can do, and how it learns.
An LLM that can answer questions about me from a JSON file is not that impressive. But it’s a platform for me to experiment from.
Notes
- Dev agents are crazy! v0 was absolutely magical. It was frustrating when something didn’t work, and I want a better mechanism for syncing with a repo, but otherwise it was great.
- I sometimes forgot that I was the expert. I found myself asking Cursor to write every single change I wanted. I sometimes would spend more time fixing a prompt than it took to just write the code.
- Models matter a lot. 4o is clearly able to understand much better than 3.5, and able to stick to instructions better. I was easily able to deceive 3.5 into providing wrong answers.
- Models don’t matter. Once I got past the initial utility of understanding, it became clear that the real juice is in data, integrations and tools.